vm_object: Modify vm_object_allocate_anon() to return OBJT_SWAP objects

[markj]

Summary

With this change, OBJT_DEFAULT objects are no longer allocated.
Instead, anonymous objects are always of type OBJT_SWAP and always have
OBJ_SWAP set.

Modify the page fault handler to check the swap block radix tree in
places where it checked for objects of type OBJT_DEFAULT. In
particular, there's no need to invoke getpages for an anonymous object
with no swap blocks assigned. It might be that we could relax this to
non-anonymous objects as well (e.g., we could perhaps avoid needlessly
calling getpages during faults on mappings of tmpfs files).

[kib]

sys/vm/vm_fault.c

235

Why do you require OBJ_ANON there?

[markj]

sys/vm/vm_fault.c

235

I think it is not strictly required. Just being conservative since the fault handler previously only looked for OBJT_DEFAULT objects, and those are almost always anonymous objects. I think I would relax this check in a separate commit, if at all.

[kib]

sys/vm/vm_fault.c

235

This is not strictly equivalent to the old condition, anyway.

I see no reason to keep OBJ_ANON.

[markj]

In _vm_object_allocate(), initialize swp.writemappings as we do the swap block
tree.

[alc]

sys/vm/vm_fault.c

227

I would replace "does" by "can" because a copy is not certain to exist even if this function returns true.

235

I found the inclusion of OBJ_ANON here to be unexpected and confusing; in other words, it forced me to think. So, I would also argue for removing it.

[markj]

Elide getpages for all swap objects with no swap blocks assigned, not just anonymous objects.

[markj]

Tweak the wording of fault_object_needs_getpages()'s comment.